The short answer
Starter reaches no DeepSeek model. Three DeepSeek identifiers are reachable on Pro, and any DeepSeek identifier outside that set resolves to Powerhouse. A free account is not model gated at all: it can try any text model, DeepSeek included, inside a lifetime allowance of 7 messages.
The DeepSeek rows carried in the Whizi cost index, with what each costs to answer once and what it costs you in credits:
| Model | Context window | Credits per message | USD per standard answer |
|---|---|---|---|
| DeepSeek V4 Flash 0731 | 1,310,720 tokens | 1 | $0.00028 |
| DeepSeek V3.2 | 164K tokens | 1 | $0.000469 |
| DeepSeek V3.1 | 164K tokens | 1 | $0.000725 |
| DeepSeek R1 | 64K tokens | 2 | $0.00195 |
| DeepSeek V4 Pro 0813 | 1M tokens | 2 | $0.00297 |
The USD column is one standard answer of 1,000 input and 500 output tokens, at OpenRouter rates read on 20 August 2026. Per thousand answers those rows run from $0.28 to $2.97.
Where the DeepSeek rows sit against the rest of the index
The cost index prices 100 rows, not the catalogue. Against that index, the DeepSeek rows land as follows.
| Index reference point | Model | USD per standard answer |
|---|---|---|
| Cheapest priced row overall | Ling-3.0-flash, inclusionAI | $0.000053 |
| Cheapest DeepSeek row | DeepSeek V4 Flash 0731 | $0.00028 |
| Median priced row | Kimi K2 Thinking | $0.00185 |
| Priciest DeepSeek row | DeepSeek V4 Pro 0813 | $0.00297 |
| Priciest priced row overall | Claude Opus 4.7 (Fast) | $0.105 |
The spread between the cheapest and the priciest priced row in the index is roughly 2000x, and every DeepSeek row sits in the lower part of that range. The largest context window anywhere in the price index belongs to a DeepSeek V4 Flash row at 1,310,720 tokens. 47 of the 100 priced rows carry a window of 1,000,000 tokens or more.
What a DeepSeek message costs in credits
A model credit cost is a fixed integer per message. Nothing about the length of your message, the length of the answer, or the length of the conversation enters the calculation, so a one line question on a DeepSeek row costs exactly what a long one does. Every DeepSeek row priced in the index sits on the first or second rung of the credit ladder. The ladder itself is in the credits reference.
What each plan allowance buys in DeepSeek messages:
| Plan | Monthly credits | DeepSeek rows reachable | Messages on a 1 credit row | Messages on a 2 credit row |
|---|---|---|---|---|
| Free | 0 credits, 7 message lifetime allowance | Any text model, inside those 7 messages | 7 in total, across all models | 7 in total, across all models |
| Starter | 400 | None | Not available | Not available |
| Pro | 2,000 | Three identifiers | 2,000 | 1,000 |
| Powerhouse | 8,000 | All of them | 8,000 | 4,000 |
That credit arithmetic is what a credit metered platform sees. Credit pricing is gated per platform and the production allowlist today is the web client. A platform that is not on the allowlist is charged one per turn and spends against the older message allowance instead, which is 400 messages a monthly period on Starter, 800 on Pro and 5,000 on Powerhouse. Message limits and credit limits coexist, and which one applies is decided per request.
Weekly billing, which is offered on mobile only, carries its own allowances: 100 credits on Starter, 500 on Pro and 2,000 on Powerhouse per weekly period. There is no grant event and no top up job behind any of this. Usage is summed per period key, so the arrival of a new period is itself the reset.
Which plan opens which DeepSeek model
Model access is resolved on the server by identifier, checked against the Starter set, then the agent set, then the Pro set, and otherwise returned as Powerhouse.
Starter includes no DeepSeek model. Starter reaches exactly four picker entries: Auto, the house model Whizi AI, one base OpenAI model and one fast Gemini model. Starter is $15.99/month, or $10.99/mo billed annually at $131.88. Pro is the lowest plan that reaches any DeepSeek model.
Pro opens three DeepSeek identifiers. Two are V3 generation rows and one is an R1 reasoning snapshot, written out by identifier in the table below. What the DeepSeek promotion into Pro selected for is the V3 generation plus a single reasoning row, not the family as a whole. Pro is $29.99/month, and it is a strict superset of Starter.
Everything else in the family needs Powerhouse. Any DeepSeek identifier that is not one of those three requires Powerhouse, which is $49.99/month. The V4 generation rows in the table above are outside the Pro list, so the newest DeepSeek rows and the largest DeepSeek window are the ones that need the top plan.
| Plan | DeepSeek identifiers it opens |
|---|---|
| Starter | None |
| Pro | deepseek deepseek-v3.2, deepseek deepseek-chat-v3.1, deepseek deepseek-r1-0528 |
| Powerhouse | Every other DeepSeek identifier in the catalogue |
A free account is the one case with no model gate. The chat path returns for the free tier before the model gate is reached, because the funnel deliberately lets prospects try any text model inside the 7 message cap.
The DeepSeek row where the token budget is clamped
Within the DeepSeek rows, only one sits under the clamp threshold. Any model whose context window sits below 93,000 tokens gets a smaller, proportional per turn budget instead of the flat one, and the R1 row at 64K is the DeepSeek row under that line. Every other DeepSeek row in the table above sits far above it.
On a 64,000 token window the clamp caps output at 40 percent of the window, which is 25,600 tokens, and holds back a 1,000 token safety margin. If you are pasting long documents into a reasoning conversation, that is the row where you will feel it first.
The unclamped rows all share one allowance. A catalogue model gets a flat budget of 40,000 input tokens and 20,000 output tokens per turn unless its window is under 93,000 tokens, so the 1,310,720 token window on the V4 Flash row and the 164K windows on the V3 rows draw the same per turn ceiling. A context window is conversation headroom, not a bigger per turn budget.
What this does not include
No DeepSeek row generates media. The image, video and audio sections of the picker are hand curated lists, and the rows in them are Whizi Image, Nano Banana, Flux and Stable Diffusion for images, Whizi Video AI for video, and Whizi Music AI for audio. None of them is a DeepSeek model.
No DeepSeek row is an agent persona. The four personas plus their three legacy aliases all run on the same base model and are priced separately from the catalogue, and personas are a Pro feature rather than a model tier.
Auto never picks DeepSeek. Auto carries six rungs: a quick rung on the house model at 1 credit, a rewrite rung and an attachment rung on a fast Gemini model at 2, a long form rung on a mid GPT model at 4, a code rung on Claude Sonnet at 10, and a reasoning rung on a top GPT model at 20. No DeepSeek row is on that ladder, so reaching DeepSeek means selecting it in the picker.
There is no per model capability flag to publish. Web search is a request level mode and attachments are handled per request, so neither is a property of a DeepSeek row, and no source records a knowledge cutoff per model.
How the DeepSeek list changes over time
Two policies shape what gets added. Nothing older than a 2025 release is added to the catalogue any more, an owner rule from 19 August 2026. And deep research identifiers are excluded on principle, because they bill a whole multi search run rather than a turn.
The picker contents are served live by the Whizi backend, so the app list is always the current one. Older DeepSeek generations stay listed alongside the newer ones because a model never leaves the catalogue once it has been offered: a conversation carries the model it was created with, so removing a row would break the threads already on it.
- Starter includes no DeepSeek model
- Pro opens three DeepSeek identifiers, Powerhouse opens the rest
- The priced DeepSeek rows cost 1 or 2 credits per message
- Credit cost is fixed per model, not per message length
- A free account can try any text model inside 7 lifetime messages
- Catalogue models run on a flat 40,000 in, 20,000 out per turn budget unless their window is under 93,000 tokens
- DeepSeek R1 at 64K is clamped below that flat budget
Frequently asked questions
How many credits does a DeepSeek message cost?
1 or 2 credits on the rows the cost index prices. The consequence worth planning around is the refusal rule: a turn that does not fit in the remaining balance is refused whole rather than part charged, so 1 credit left is not enough to send a 2 credit DeepSeek turn. Running out returns HTTP 429 with the code message_limit_reached.
Can I use DeepSeek on the Starter plan?
No. The gate runs on the server by model identifier rather than in the interface, so a locked row cannot be reached by editing a request: selecting one returns HTTP 403 with the code tier_upgrade_required and the message "Upgrade your plan to use this model."
Which DeepSeek model has the largest context window in Whizi?
The DeepSeek V4 Flash row, at 1,310,720 tokens, which is the largest window of any row in the Whizi price index. It resolves to Powerhouse, so the largest window in the index is not one of the rows a Pro subscription opens.
Does a DeepSeek reasoning answer cost more than a short one?
No. What a long answer costs you is room, not credits: the R1 row carries the smallest window in the family at 64K, which puts it under the 93,000 token clamp threshold, so its per turn output is capped at 25,600 tokens with a 1,000 token margin held back while the V3 and V4 rows draw the full flat budget. Moving that thread onto a V3 or V4 row is a per message decision, and each message is charged at the rung of whichever row answered it.
Can DeepSeek read a PDF in Whizi?
In the web app, PDF, Word and spreadsheet text is extracted in your browser and sent as text to whichever model you selected, so a DeepSeek row receives text rather than a binary file. That extracted text spends the same per turn input budget as anything else you type, so what bounds a long document is the per turn ceiling rather than the size of the file you dropped in. The accepted formats and how each one is handled are in supported file types.